Unsupervised Class-Expert Learning for Supporting Covid-19 Triage Based on Computed Tomography Data

Abstract

Deep learning applications in medical imaging have been achieving promising results in the detection of diseases, among which clinical trials in terms of screening and diagnosis of patients with COVID-19 stand out. Computed Tomography (CT) images of the chest have been used by specialists for the diagnosis of COVID-19. However, due to the need of the moment and the possibility of using computational resources to help the medical team, it is observed in the literature several proposed works using supervised learning, however it lacks unsupervised methods for the screening and diagnosis of patients with COVID-19. In this work, the deep learning models Convolutional Neural Network (CNN) and Variational Autoencoders are used for feature extraction and later this information is used for binary and multiclass classification in unsupervised methods (k-means, Fuzzy C-Means and Self-Organizing Maps). For this purpose, a public database containing 4173 CT images (2168 CT slices from COVID-19, 758 slices from Healthy and 1247 slices from other lung diseases) was used. The results show that feature extraction via Variational Autoencoders has similar performance with state-of-the-art models in the literature for COVID-19, mainly for the binary classification with accuracies of 95.9%, 92.1% and 95.9% for k-means, Fuzzy C-Means and SOM, respectively, presenting competitive results in the literature. It also shows the importance of extracting features through convolutional networks to improve classification performance, resulting from the use of deep learning and its state of the art in the area of computer vision.